One knob,
a whole self.
“NA level is arousal. The LC-NA system is the model’s arousal mechanism. There is no separate ‘arousal’ abstraction that then derives NA — that is circular.”
— taxonomy-below-situations.md, design principle 4
Drag a self from coma to optimal to collapse and eighty cortical regions rearrange as you go, all of it driven by one brainstem nucleus working two dials. Every readout is either a lookup into a real six-thousand-point simulation or a live pass through the model's own gain equation.
ENTER THE INSTRUMENTThe heartbeat
One oscillator sits under everything downstream. The Locus Coeruleus (a nucleus the size of a grain of rice) is modelled as a single FitzHugh-Nagumo unit: a firing-rate proxy v chased by a recovery variable w. Feed it a gentle current and it settles into a steadytonic murmur; drive it harder and the same unit snaps into sharpphasic bursts. Only the positive excursions release noradrenaline into the cortex.
a = 0.7 · b = 0.8 · τ = 12.5 · Itonic = 0.65 · Iphasic = 1.5src/models/lc.py
The one knob
The dashed line is the theory, the real inverted-U gain function (Arnsten 2011). The orange trace is what the model actually produced: the sweep’s own mean excitation, normalised to fill the axis, so it climbs almost monotonically and refuses to fall, ignoring the inverted-U it is meant to follow (the raw Hz barely move at all; Act V shows how little). Move the playhead up the arousal spine, with a pointer, the ← → arrow keys or the slider below, and the cortex, the qualia and the caption re-render from the model.
THE CORTEX FIELD
80 AAL2 regions · brightness = each hemisphere's inverted-U gain(NA) × its NET densityLC-DRIVEN MODEL LOADS
nc-model bridge · loads > 0.30An illustrative bridge between the two models on this page. The developmental cascade model (Act VIII) runs with neurotypical defaults everywhere except the shared LC / NE parameters (each hemisphere's NE tone and the tonic–phasic mode), which are driven live from the arousal state above. None of this is a reading about a real person, since a momentary arousal state carries no developmental history with it. It only shows what the cascade does when its LC / NE inputs alone are perturbed and everything else is held neutral: at rest it reproduces the pure neurotypical-default output.
awaiting the parametric model…
THE FIGURE
one self, drawn from the seven readouts · re-embodies liveA single self, re-drawn from the seven qualia as you move the dials.
Loading the sweep…
EMERGENT QUALIA
7 readouts · L vs R hemisphere · normalised 0–1The two columns differ even at identical noradrenaline: the drive is symmetric, but each hemisphere reads its own connectome. Arousal runs right-dominant here, which matches the right hemisphere's known lead in global arousal and vigilance, while sustained attention and network stability lean left.
Magnets on the curve
Click a preset and the playhead snaps straight to where that state sits on the inverted-U. A drug works differently: click it and the playhead slides by the width of a real pharmacological mechanism instead of jumping. Both are the model’s illustrative hypotheses and not clinical fact. The receptor geography under them is measured, but where each landmark lands was set by hand.
STATES OF ONE MIND situational contexts · presets.py
CLINICAL PROFILES LC-NA dysfunction · presets.py · illustrative
PHARMACOLOGY each agent nudges NA along the axis · drugs.py
The whole space
The playhead lives inside a plane. This is a 41×41 downsample of the full closed-loop sweep (NA against tonic-phasic balance) from an 81×81, 6561-point neurolib run. Every value is min-max normalised per metric, because the raw field is nearly flat. Hover to read a cell; click to send the instrument there.
The flatness of a modelled mind
Sweep the whole plane and the mean firing rate barely moves. Mean cortical excitation ranges only from 18.25 to 20.40 Hz across all 6561 states. Below NA ≈ 0.15 the model returns 18.273 Hz exactly, the same number for every setting of the second dial. Of the full sweep 1,406 states fall in that floor bin, and 1,022 of them sit pinned at precisely 18.273 Hz.
A mind flattened to two numbers is mostly a plateau with a couple of cliffs. The structure you feel in the instrument lives almost entirely in the normalisation and in the seven derived qualia: the rawselective-attention channel spans 0 to 4.2 while the channel called arousal barely twitches. A model built to be about arousal says almost nothing with the quantity it calls arousal.
“Receptor geography is empirical (Hansen et al. 2022 PET data). The mixing weights … and the overall scaling coefficients … are hand-tuned parameters chosen for numerical stability. Situation profiles are illustrative defaults, not literature-derived values.”— taxonomy-below-situations.md, principle 7
This piece is algorithmic frottage, a mind rubbing against a statistical model of itself, keeping every crease the model admits to.
The geography
“Every term has receptor geography. No uniform scalars applied to all 80 regions.”— taxonomy-below-situations.md, design principle 1
What stops the one dial from being a flat multiplier is the map underneath it. Nineteen PET receptor densities from Hansen et al. (2022), measured across 1,238 people, give every region its own sensitivity. Prefrontal cortex is rich in the noradrenaline transporter and moves hard when the dial turns; primary visual cortex barely notices. This map is the one part of the model that was measured rather than tuned.
| System | Receptors & transporters | Role in the model |
|---|---|---|
| Norepinephrine | NET | the arousal transporter — proxy for LC innervation density |
| Dopamine | D1 · D2 · DAT | PFC executive drive — inverted-U on D1 (Gs), monotonic gate on D2 (Gi) |
| Serotonin | 5-HT1A · 5-HT1B · 5-HT2A · 5-HT4 · 5-HT6 · 5-HTT | affect and stress — limbic inhibition, sensory excitation |
| Acetylcholine | α4β2 · M1 · VAChT | cortical engagement — pro-attentional gain |
| Glutamate | mGluR5 · NMDA | the excitatory backbone |
| GABA | GABAa | inhibitory stability — the E/I brake |
| Histamine | H3 | wakefulness modulation |
| Cannabinoid | CB1 | network-stability term |
| Opioid | MOR | affective weighting |
The model is cortical only, 80 AAL2 regions grouped into the seven Yeo networks, with no subcortex to turn: no thalamus, striatum or amygdala.
Grounded and hand-tuned
The most honest document in the repository is an inventory of its own soft spots, roughly 120 named constants sorted by how much of the model’s character they invent. The geography is measured, but almost everything that turns it into a feeling was set by hand for the numbers to behave.
I’ve half-built a plan to fit these constants against real neuroimaging data and let the optimiser tell me which of my hand-tuned guesses were wrong. Until then the model is what it admits to being: an empirical map carrying a stack of plausible coefficients and one very confident dial. The work is in dragging a mind across that honestly, guesses and all.
Sliders → predicted diagnoses
A second model rides beneath the first. Where the instrument above turns one arousal dial, this one exposes the whole developmental cascade: fifty-one free parameters (genes, neural-crest counts, neurotransmitter tone, environment, medication) feeding a fitted network of 157 DSM criteria into seventeen diagnosis loads. Every weight was learned against the ABIDE and ADHD-200 cohorts. The whole computation is ported to run live in your browser; move a slider and the diagnosis bars re-solve from ~235 formulas.
DEVELOPMENTAL PARAMETERS
PREDICTED DIAGNOSIS LOADS
17 outputs · bars redden past 0.50DERIVED SIGNALS
computed sliders · live from the cascadeThis is a research parametric model rather than a diagnostic tool. The bars reproduce the model’s own outputs faithfully in JavaScript, not a clinical judgement about any person.
The cohort
None of those weights were invented. The parametric model was fitted to real, publicly labelled neuroimaging cohorts: the Autism Brain Imaging Data Exchange and the ADHD-200 sample. Below, the fit is tested on subjects it never trained on.
TRAINING RUNS 30 fits · AUROC and severity-r on the held-out set
REAL SUBJECTS 600 sampled ABIDE subjects · model load vs clinical severity
FITTED SUBTYPES four data-driven clusters · deviation from the population mean
Honest limits: the held-out validation set is small (≈250–274 subjects), the strongest AUROC and severity-r cannot both be claimed by a single run and the labels are research diagnoses, not clinical adjudication. It is a strong-but-not-clinical parametric fit: an illustration that a mechanistic developmental cascade can separate real cohorts, not a screening instrument.— nc-model-ml · ABIDE I/II + ADHD-200 · n≈2424